Contribution
Integrating Neural Networks with Boundary Element Methods for Robust Sound Field Calculations
* Presenting author
Abstract:
Computational methods that solve the Helmholtz equation via discretization are essential in acoustic analysis. The boundary element method is one of the conventional techniques for predicting sound radiation and wave propagation in bounded and unbounded domains. It is well suited for exterior problems as it only requires discretization of the domain boundary. Furthermore, its formulation implicitly satisfies the far-field radiation condition eliminating the need for domain truncation. However, the performance of boundary element approaches and other discretization-based methods usually deteriorates when using imperfect data as input for the computational model. Imposing boundary conditions that are subjected to noise, such as those derived from surface vibration measurements, heavily impacts the accuracy and robustness of the solution. Recently, advances in physics-informed machine learning have shown regularization capabilities when dealing with noisy input data. This study integrates neural networks with the boundary element method for robust sound field calculations. Embedding the residual of the boundary integral equation into the loss function enables data-driven predictions of acoustic fields in two-dimensional domains with noisy boundary conditions. Results indicate that boundary integral neural networks perform well for both interior and exterior problems, outperforming the traditional boundary element method in high-noise scenarios.